Samuel Thibault
|
e71124580f
shut up warning
|
8 anos atrás |
Samuel Thibault
|
91f985c664
Do not bother telling _starpu_post_data_request which node is handler for the request, it can determine it by itself from the request, and it is actually more right when the request is an invalidation, for which the handler is undetermined (-1)
|
8 anos atrás |
Corentin Salingue
|
cc4eb6b654
fix indent
|
8 anos atrás |
Samuel Thibault
|
0ae8f154db
fix warning
|
8 anos atrás |
Corentin Salingue
|
087b676559
add MPI slave to MPI slave sync transfers
|
8 anos atrás |
Corentin Salingue
|
0dbf966aef
Merge from trunk
|
8 anos atrás |
Corentin Salingue
|
f7acc7fe76
* fix bug in async mode when using MPI Master-Slave
|
8 anos atrás |
Corentin Salingue
|
3bbc1443c4
Each worker can manage the good set of memories
|
8 anos atrás |
Nathalie Furmento
|
ea326f4d15
src: add character in front of macros STARPU_MALLOC, STARPU_CALLOC and STARPU_REALLOC
|
8 anos atrás |
Nathalie Furmento
|
d4761fddf6
src: introduce macros STARPU_MALLOC, STARPU_CALLOC and STARPU_REALLOC to make sure we always check the return value of stdlib malloc functions
|
8 anos atrás |
Samuel Thibault
|
b643914328
fix warnings
|
8 anos atrás |
Samuel Thibault
|
c0b04e4d0d
Fix and test the case where we try to invalidate a data which has some transfers pending. Also make sure that eviction doesn't evict a data being invalidated
|
8 anos atrás |
Samuel Thibault
|
0b5135e60e
re-apply r18403 again, we'll be using maxmin/precision simgrid parameter to coarsen the time
|
8 anos atrás |
Samuel Thibault
|
514f24d352
revert r18403 for now, it seems to increase starpu-mpi-simgrid execution time a lot
|
9 anos atrás |
Samuel Thibault
|
4fb4ce92aa
Reduce the granularity of simulation of CUDA and MPI probes, it is not that expensive.
|
9 anos atrás |
Samuel Thibault
|
6bd3bb1b96
revert spurious change
|
9 anos atrás |
Samuel Thibault
|
8bb9a66912
Fix data migration: we need to flush the cache too. Add starpu_mpi_data_migrate which does things properly
|
9 anos atrás |
Samuel Thibault
|
1f841a3f0d
Add origin of data requests for debugging
|
9 anos atrás |
Samuel Thibault
|
38097a0048
fix spurious trace event
|
9 anos atrás |
Samuel Thibault
|
b75b95b108
fix printing data requests in gdb
|
9 anos atrás |
Samuel Thibault
|
a5f7d96fea
Add rbtree and rbtree+list implementation. Use it to introduce priorities in data requests, which default to the task priority or workerorder
|
9 anos atrás |
Samuel Thibault
|
e56836b055
Keep fetches and prefetches in the submission order even when upgrading
|
9 anos atrás |
Samuel Thibault
|
0e92653f70
Add _list_member and _list_ismember helpers, and use them
|
9 anos atrás |
Samuel Thibault
|
ecc274e4c0
Distinguish between fetches, prefetches and idle fetches in the paje trace
|
9 anos atrás |
Samuel Thibault
|
90bf39960e
Vastly improve simgrid simulation time by using wake queues
|
9 anos atrás |
Samuel Thibault
|
23b5457f92
port r16114 from 1.2: Make non-cpu drivers also process main_ram requests so write_invalidation requests can work in non-cpu case. Those drivers can run them anyway, and this is actually required for the disk+gpu case.
|
9 anos atrás |
Samuel Thibault
|
cef92ef39b
Catch bogus handling_node early
|
9 anos atrás |
Samuel Thibault
|
91fd2cfc15
Fix mem_reclaim: on write request, if some requests are pending (notably reclaiming), create an additional write request which will wait for them before invalidating all copies
|
9 anos atrás |
Samuel Thibault
|
d06eb3bbd4
Make reclaiming prevent more fetches on the node being cleaned up, to avoid permanent battle between them
|
10 anos atrás |
Samuel Thibault
|
93e54d53ad
Fix safety of _starpu_wait_data_request_completion: starpu_data_request_destroy was supposed to be called with the handle lock held. Move that part of starpu_data_request_destroy to _starpu_data_request_unlink, which can then be called directly from _starpu_wait_data_request_completion, avoiding races between unregistering the handle and _starpu_wait_data_request_completion
|
10 anos atrás |